Datawhale 零基礎入門數據挖掘-Task4 建模調參
四、建模與調參
Tip:此部分爲零基礎入門數據挖掘的 Task4 建模調參 部分,帶你來了解各種模型以及模型的評價和調參策略,歡迎大家後續多多交流。
賽題:零基礎入門數據挖掘 - 二手車交易價格預測
地址:https://tianchi.aliyun.com/competition/entrance/231784/introduction?spm=5176.12281957.1004.1.38b02448ausjSX
5.1 學習目標
- 瞭解常用的機器學習模型,並掌握機器學習模型的建模與調參流程
- 完成相應學習打卡任務
5.2 內容介紹
- 線性迴歸模型:
- 線性迴歸對於特徵的要求;
- 處理長尾分佈;
- 理解線性迴歸模型;
- 模型性能驗證:
- 評價函數與目標函數;
- 交叉驗證方法;
- 留一驗證方法;
- 針對時間序列問題的驗證;
- 繪製學習率曲線;
- 繪製驗證曲線;
- 嵌入式特徵選擇:
- Lasso迴歸;
- Ridge迴歸;
- 決策樹;
- 模型對比:
- 常用線性模型;
- 常用非線性模型;
- 模型調參:
- 貪心調參方法;
- 網格調參方法;
- 貝葉斯調參方法;
5.3 相關原理介紹與推薦
由於相關算法原理篇幅較長,本文推薦了一些博客與教材供初學者們進行學習。
5.3.1 線性迴歸模型
https://zhuanlan.zhihu.com/p/49480391
5.3.2 決策樹模型
https://zhuanlan.zhihu.com/p/65304798
5.3.3 GBDT模型
https://zhuanlan.zhihu.com/p/45145899
5.3.4 XGBoost模型
https://zhuanlan.zhihu.com/p/86816771
5.3.5 LightGBM模型
https://zhuanlan.zhihu.com/p/89360721
5.3.6 推薦教材:
- 《機器學習》 https://book.douban.com/subject/26708119/
- 《統計學習方法》 https://book.douban.com/subject/10590856/
- 《Python大戰機器學習》 https://book.douban.com/subject/26987890/
- 《面向機器學習的特徵工程》 https://book.douban.com/subject/26826639/
- 《數據科學家訪談錄》 https://book.douban.com/subject/30129410/
5.4 代碼示例
5.4.1 讀取數據
import pandas as pd
import numpy as np
import warnings
warnings.filterwarnings('ignore')
reduce_mem_usage 函數通過調整數據類型,幫助我們減少數據在內存中佔用的空間
def reduce_mem_usage(df):
""" iterate through all the columns of a dataframe and modify the data type
to reduce memory usage.
"""
start_mem = df.memory_usage().sum()
print('Memory usage of dataframe is {:.2f} MB'.format(start_mem))
for col in df.columns:
col_type = df[col].dtype
if col_type != object:
c_min = df[col].min()
c_max = df[col].max()
if str(col_type)[:3] == 'int':
if c_min > np.iinfo(np.int8).min and c_max < np.iinfo(np.int8).max:
df[col] = df[col].astype(np.int8)
elif c_min > np.iinfo(np.int16).min and c_max < np.iinfo(np.int16).max:
df[col] = df[col].astype(np.int16)
elif c_min > np.iinfo(np.int32).min and c_max < np.iinfo(np.int32).max:
df[col] = df[col].astype(np.int32)
elif c_min > np.iinfo(np.int64).min and c_max < np.iinfo(np.int64).max:
df[col] = df[col].astype(np.int64)
else:
if c_min > np.finfo(np.float16).min and c_max < np.finfo(np.float16).max:
df[col] = df[col].astype(np.float16)
elif c_min > np.finfo(np.float32).min and c_max < np.finfo(np.float32).max:
df[col] = df[col].astype(np.float32)
else:
df[col] = df[col].astype(np.float64)
else:
df[col] = df[col].astype('category')
end_mem = df.memory_usage().sum()
print('Memory usage after optimization is: {:.2f} MB'.format(end_mem))
print('Decreased by {:.1f}%'.format(100 * (start_mem - end_mem) / start_mem))
return df
sample_feature = reduce_mem_usage(pd.read_csv('data_for_tree.csv'))
Memory usage of dataframe is 57322784.00 MB
Memory usage after optimization is: 14755178.00 MB
Decreased by 74.3%
sample_feature.head()
name | model | brand | bodyType | fuelType | gearbox | power | kilometer | notRepairedDamage | price | ... | used_time | city | brand_amount | brand_price_max | brand_price_median | brand_price_min | brand_price_sum | brand_price_std | brand_price_average | power_bin | |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | 736 | 30 | 6 | 1.0 | 0.0 | 0.0 | 60 | 12.5 | 0.0 | 1850.0 | ... | 4384.0 | 1.0 | 10192.0 | 35990.0 | 1800.0 | 13.0 | 36457520.0 | 4564.0 | 3576.0 | 5.0 |
1 | 2262 | 40 | 1 | 2.0 | 0.0 | 0.0 | 0 | 15.0 | MISSING | 3600.0 | ... | 4756.0 | 4.0 | 13656.0 | 84000.0 | 6400.0 | 15.0 | 124044600.0 | 8992.0 | 9080.0 | NaN |
2 | 14874 | 115 | 15 | 1.0 | 0.0 | 0.0 | 163 | 12.5 | 0.0 | 6222.0 | ... | 4384.0 | 2.0 | 1458.0 | 45000.0 | 8496.0 | 100.0 | 14373814.0 | 5424.0 | 9848.0 | 16.0 |
3 | 71865 | 109 | 10 | 0.0 | 0.0 | 1.0 | 193 | 15.0 | 0.0 | 2400.0 | ... | 7124.0 | NaN | 13992.0 | 92900.0 | 5200.0 | 15.0 | 113034208.0 | 8248.0 | 8076.0 | 19.0 |
4 | 111080 | 110 | 5 | 1.0 | 0.0 | 0.0 | 68 | 5.0 | 0.0 | 5200.0 | ... | 1531.0 | 6.0 | 4664.0 | 31500.0 | 2300.0 | 20.0 | 15414322.0 | 3344.0 | 3306.0 | 6.0 |
5 rows × 36 columns
continuous_feature_names = [x for x in sample_feature.columns if x not in ['price','brand','model','name', 'bodyType', 'fuelType', 'notRepairedDamage']]
continuous_feature_names
['gearbox',
'power',
'kilometer',
'v_0',
'v_1',
'v_2',
'v_3',
'v_4',
'v_5',
'v_6',
'v_7',
'v_8',
'v_9',
'v_10',
'v_11',
'v_12',
'v_13',
'v_14',
'train',
'used_time',
'city',
'brand_amount',
'brand_price_max',
'brand_price_median',
'brand_price_min',
'brand_price_sum',
'brand_price_std',
'brand_price_average',
'power_bin']
5.4.2 線性迴歸 & 五折交叉驗證 & 模擬真實業務情況
sample_feature = sample_feature.dropna().replace('-', 0).reset_index(drop=True)
sample_feature = sample_feature.replace('MISSING', 0)
print(sample_feature.head())
sample_feature['notRepairedDamage'] = sample_feature['notRepairedDamage'].astype(np.float32)
train = sample_feature[continuous_feature_names + ['price']]
train_X = train[continuous_feature_names]
train_y = train['price']
name model brand bodyType fuelType gearbox power kilometer \
0 736 30 6 1.0 0.0 0.0 60 12.5
1 14874 115 15 1.0 0.0 0.0 163 12.5
2 111080 110 5 1.0 0.0 0.0 68 5.0
3 137642 24 10 0.0 1.0 0.0 109 10.0
4 2402 13 4 0.0 0.0 1.0 150 15.0
notRepairedDamage price ... used_time city brand_amount \
0 0.0 1850.0 ... 4384.0 1.0 10192.0
1 0.0 6222.0 ... 4384.0 2.0 1458.0
2 0.0 5200.0 ... 1531.0 6.0 4664.0
3 0.0 8000.0 ... 2482.0 3.0 13992.0
4 0.0 3500.0 ... 6184.0 3.0 16576.0
brand_price_max brand_price_median brand_price_min brand_price_sum \
0 35990.0 1800.0 13.0 36457520.0
1 45000.0 8496.0 100.0 14373814.0
2 31500.0 2300.0 20.0 15414322.0
3 92900.0 5200.0 15.0 113034208.0
4 99999.0 6000.0 12.0 138279072.0
brand_price_std brand_price_average power_bin
0 4564.0 3576.0 5.0
1 5424.0 9848.0 16.0
2 3344.0 3306.0 6.0
3 8248.0 8076.0 10.0
4 8088.0 8344.0 14.0
[5 rows x 36 columns]
5.4.2 - 1 簡單建模
from sklearn.linear_model import LinearRegression
model = LinearRegression(normalize=True)
model = model.fit(train_X, train_y)
查看訓練的線性迴歸模型的截距(intercept)與權重(coef)
print('intercept:'+ str(model.intercept_))
sorted(dict(zip(continuous_feature_names, model.coef_)).items(), key=lambda x:x[1], reverse=True)
intercept:-763121.415400561
[('v_6', 3418891.8661930403),
('v_5', 2297424.5932109547),
('v_8', 1219155.1076147154),
('v_9', 689318.2798646946),
('v_7', 441213.76855152246),
('v_11', 33695.21532491009),
('v_12', 12751.353677724243),
('v_10', 11940.051759841095),
('gearbox', 915.6123144165624),
('v_14', 232.73304565903075),
('city', 38.480371412194856),
('power', 35.6062576498682),
('brand_price_median', 0.45503872387225264),
('brand_price_std', 0.45190937612308135),
('brand_amount', 0.18847929674706093),
('brand_price_max', 0.005052224552260477),
('train', 3.818422555923462e-07),
('brand_price_sum', -3.0058302585516643e-05),
('used_time', -0.02462480273046145),
('brand_price_average', -0.34643690063272586),
('brand_price_min', -2.5319297835051704),
('power_bin', -109.76386262213285),
('v_13', -150.7399294462697),
('kilometer', -351.36103146189294),
('v_3', -1344.0675611153686),
('v_0', -2798.561647500838),
('v_4', -3853.612550388567),
('v_2', -39428.5597862382),
('v_1', -44599.82891284008)]
from matplotlib import pyplot as plt
subsample_index = np.random.randint(low=0, high=len(train_y), size=50)
繪製特徵v_9的值與標籤的散點圖,圖片發現模型的預測結果(藍色點)與真實標籤(黑色點)的分佈差異較大,且部分預測值出現了小於0的情況,說明我們的模型存在一些問題
plt.scatter(train_X['v_9'][subsample_index], train_y[subsample_index], color='black')
plt.scatter(train_X['v_9'][subsample_index], model.predict(train_X.loc[subsample_index]), color='blue')
plt.xlabel('v_9')
plt.ylabel('price')
plt.legend(['True Price','Predicted Price'],loc='upper right')
print('The predicted price is obvious different from true price')
plt.show()
The predicted price is obvious different from true price
通過作圖我們發現數據的標籤(price)呈現長尾分佈,不利於我們的建模預測。原因是很多模型都假設數據誤差項符合正態分佈,而長尾分佈的數據違背了這一假設。參考博客:https://blog.csdn.net/Noob_daniel/article/details/76087829
import seaborn as sns
print('It is clear to see the price shows a typical exponential distribution')
plt.figure(figsize=(15,5))
plt.subplot(1,2,1)
sns.distplot(train_y)
plt.subplot(1,2,2)
sns.distplot(train_y[train_y < np.quantile(train_y, 0.9)])
It is clear to see the price shows a typical exponential distribution
<matplotlib.axes._subplots.AxesSubplot at 0x1833c28a4c8>
在這裏我們對標籤進行了 變換,使標籤貼近於正態分佈
train_y_ln = np.log(train_y + 1)
import seaborn as sns
print('The transformed price seems like normal distribution')
plt.figure(figsize=(15,5))
plt.subplot(1,2,1)
sns.distplot(train_y_ln)
plt.subplot(1,2,2)
sns.distplot(train_y_ln[train_y_ln < np.quantile(train_y_ln, 0.9)])
The transformed price seems like normal distribution
<matplotlib.axes._subplots.AxesSubplot at 0x1833c60cd08>
model = model.fit(train_X, train_y_ln)
print('intercept:'+ str(model.intercept_))
sorted(dict(zip(continuous_feature_names, model.coef_)).items(), key=lambda x:x[1], reverse=True)
intercept:22.490527976445637
[('v_9', 6.73837719656898),
('v_1', 1.8764010743138013),
('v_12', 1.5490066205584243),
('v_5', 1.3684828986478352),
('v_13', 0.9381007016475442),
('v_11', 0.8601076136541934),
('v_3', 0.6908662876168846),
('v_7', 0.07176605184338732),
('power_bin', 0.009208120503260045),
('gearbox', 0.005832463904491905),
('power', 0.0004532988300577831),
('brand_price_min', 2.958178448217908e-05),
('used_time', 7.25708186886524e-06),
('brand_amount', 3.2317294039114087e-06),
('brand_price_median', 1.2316687308102237e-06),
('brand_price_max', 7.440945604426392e-07),
('brand_price_average', 6.077520449532623e-07),
('train', -1.2789769243681803e-11),
('brand_price_sum', -2.437300649289914e-10),
('brand_price_std', -4.2133697033978156e-07),
('v_14', -0.0003021985128370997),
('city', -0.003247381687803047),
('kilometer', -0.012962886393843829),
('v_0', -0.031397921158372956),
('v_2', -0.698190677552077),
('v_4', -0.8159958185074844),
('v_10', -1.5348138603344603),
('v_8', -42.38488913963534),
('v_6', -253.24942729281895)]
再次進行可視化,發現預測結果與真實值較爲接近,且未出現異常狀況
plt.scatter(train_X['v_9'][subsample_index], train_y[subsample_index], color='black')
plt.scatter(train_X['v_9'][subsample_index], np.exp(model.predict(train_X.loc[subsample_index])), color='blue')
plt.xlabel('v_9')
plt.ylabel('price')
plt.legend(['True Price','Predicted Price'],loc='upper right')
print('The predicted price seems normal after np.log transforming')
plt.show()
The predicted price seems normal after np.log transforming
5.4.2 - 2 五折交叉驗證
K折交叉驗證
由於驗證數據集不參與模型訓練,當訓練數據不夠用時,預留大量的驗證數據顯得太奢侈。一種改善的方法是K折交叉驗證(K-fold cross-validation)。在K折交叉驗證中,我們把原始訓練數據集分割成K個不重合的子數據集,然後我們做K次模型訓練和驗證。每一次,我們使用一個子數據集驗證模型,並使用其他K-1個子數據集來訓練模型。在這K次訓練和驗證中,每次用來驗證模型的子數據集都不同。最後,我們對這K次訓練誤差和驗證誤差分別求平均。
在使用訓練集對參數進行訓練的時候,經常會發現人們通常會將一整個訓練集分爲三個部分(比如mnist手寫訓練集)。一般分爲:訓練集(train_set),評估集(valid_set),測試集(test_set)這三個部分。這其實是爲了保證訓練效果而特意設置的。其中測試集很好理解,其實就是完全不參與訓練的數據,僅僅用來觀測測試效果的數據。而訓練集和評估集則牽涉到下面的知識了。
因爲在實際的訓練中,訓練的結果對於訓練集的擬合程度通常還是挺好的(初始條件敏感),但是對於訓練集之外的數據的擬合程度通常就不那麼令人滿意了。因此我們通常並不會把所有的數據集都拿來訓練,而是分出一部分來(這一部分不參加訓練)對訓練集生成的參數進行測試,相對客觀的判斷這些參數對訓練集之外的數據的符合程度。這種思想就稱爲交叉驗證(Cross Validation)
from sklearn.model_selection import cross_val_score
from sklearn.metrics import mean_absolute_error, make_scorer ## MAE
def log_transfer(func):
def wrapper(y, yhat):
result = func(np.log(y), np.nan_to_num(np.log(yhat)))
return result
return wrapper
scores = cross_val_score(model, X=train_X, y=train_y, verbose=1, cv = 5, scoring=make_scorer(log_transfer(mean_absolute_error)))
[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
[Parallel(n_jobs=1)]: Done 5 out of 5 | elapsed: 1.6s finished
使用線性迴歸模型,對未處理標籤的特徵數據進行五折交叉驗證(Error 1.36)
print('AVG:', np.mean(scores))
AVG: 1.369013918691876
使用線性迴歸模型,對處理過標籤的特徵數據進行五折交叉驗證(Error 0.19)
scores = cross_val_score(model, X=train_X, y=train_y_ln, verbose=1, cv = 5, scoring=make_scorer(mean_absolute_error))
[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
[Parallel(n_jobs=1)]: Done 5 out of 5 | elapsed: 1.5s finished
print('AVG:', np.mean(scores))
AVG: 0.19576915594425995
scores = pd.DataFrame(scores.reshape(1,-1))
scores.columns = ['cv' + str(x) for x in range(1, 6)]
scores.index = ['MAE']
scores
cv1 | cv2 | cv3 | cv4 | cv5 | |
---|---|---|---|---|---|
MAE | 0.194274 | 0.195956 | 0.195945 | 0.194693 | 0.197977 |
5.4.2 - 3 模擬真實業務情況
但在事實上,由於我們並不具有預知未來的能力,五折交叉驗證在某些與時間相關的數據集上反而反映了不真實的情況。通過2018年的二手車價格預測2017年的二手車價格,這顯然是不合理的,因此我們還可以採用時間順序對數據集進行分隔。在本例中,我們選用靠前時間的4/5樣本當作訓練集,靠後時間的1/5當作驗證集,最終結果與五折交叉驗證差距不大
import datetime
sample_feature = sample_feature.reset_index(drop=True)
sample_feature
name | model | brand | bodyType | fuelType | gearbox | power | kilometer | notRepairedDamage | price | ... | used_time | city | brand_amount | brand_price_max | brand_price_median | brand_price_min | brand_price_sum | brand_price_std | brand_price_average | power_bin | |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | 736 | 30 | 6 | 1.0 | 0.0 | 0.0 | 60 | 12.5 | 0.0 | 1850.0 | ... | 4384.0 | 1.0 | 10192.0 | 35990.0 | 1800.0 | 13.0 | 36457520.0 | 4564.0 | 3576.0 | 5.0 |
1 | 14874 | 115 | 15 | 1.0 | 0.0 | 0.0 | 163 | 12.5 | 0.0 | 6222.0 | ... | 4384.0 | 2.0 | 1458.0 | 45000.0 | 8496.0 | 100.0 | 14373814.0 | 5424.0 | 9848.0 | 16.0 |
2 | 111080 | 110 | 5 | 1.0 | 0.0 | 0.0 | 68 | 5.0 | 0.0 | 5200.0 | ... | 1531.0 | 6.0 | 4664.0 | 31500.0 | 2300.0 | 20.0 | 15414322.0 | 3344.0 | 3306.0 | 6.0 |
3 | 137642 | 24 | 10 | 0.0 | 1.0 | 0.0 | 109 | 10.0 | 0.0 | 8000.0 | ... | 2482.0 | 3.0 | 13992.0 | 92900.0 | 5200.0 | 15.0 | 113034208.0 | 8248.0 | 8076.0 | 10.0 |
4 | 2402 | 13 | 4 | 0.0 | 0.0 | 1.0 | 150 | 15.0 | 0.0 | 3500.0 | ... | 6184.0 | 3.0 | 16576.0 | 99999.0 | 6000.0 | 12.0 | 138279072.0 | 8088.0 | 8344.0 | 14.0 |
... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |
96262 | 43073 | 42 | 1 | 1.0 | 0.0 | 0.0 | 122 | 3.0 | 0.0 | 14780.0 | ... | 1538.0 | 5.0 | 13656.0 | 84000.0 | 6400.0 | 15.0 | 124044600.0 | 8992.0 | 9080.0 | 12.0 |
96263 | 163978 | 121 | 10 | 4.0 | 0.0 | 1.0 | 163 | 15.0 | 0.0 | 5900.0 | ... | 5772.0 | 4.0 | 13992.0 | 92900.0 | 5200.0 | 15.0 | 113034208.0 | 8248.0 | 8076.0 | 16.0 |
96264 | 184535 | 116 | 11 | 0.0 | 0.0 | 0.0 | 125 | 10.0 | 0.0 | 9500.0 | ... | 2322.0 | 2.0 | 2944.0 | 34500.0 | 2900.0 | 30.0 | 13398006.0 | 4724.0 | 4548.0 | 12.0 |
96265 | 147587 | 60 | 11 | 1.0 | 1.0 | 0.0 | 90 | 6.0 | 0.0 | 7500.0 | ... | 2003.0 | 3.0 | 2944.0 | 34500.0 | 2900.0 | 30.0 | 13398006.0 | 4724.0 | 4548.0 | 8.0 |
96266 | 45907 | 34 | 10 | 3.0 | 1.0 | 0.0 | 156 | 15.0 | 0.0 | 4999.0 | ... | 3672.0 | 1.0 | 13992.0 | 92900.0 | 5200.0 | 15.0 | 113034208.0 | 8248.0 | 8076.0 | 15.0 |
96267 rows × 36 columns
split_point = len(sample_feature) // 5 * 4
train = sample_feature.loc[:split_point].dropna()
val = sample_feature.loc[split_point:].dropna()
train_X = train[continuous_feature_names]
train_y_ln = np.log(train['price'] + 1)
val_X = val[continuous_feature_names]
val_y_ln = np.log(val['price'] + 1)
model = model.fit(train_X, train_y_ln)
mean_absolute_error(val_y_ln, model.predict(val_X))
0.19796660363310997
5.4.2 - 4 繪製學習率曲線與驗證曲線
from sklearn.model_selection import learning_curve, validation_curve
? learning_curve
def plot_learning_curve(estimator, title, X, y, ylim=None, cv=None,n_jobs=1, train_size=np.linspace(.1, 1.0, 5 )):
plt.figure()
plt.title(title)
if ylim is not None:
plt.ylim(*ylim)
plt.xlabel('Training example')
plt.ylabel('score')
train_sizes, train_scores, test_scores = learning_curve(estimator, X, y, cv=cv, n_jobs=n_jobs, train_sizes=train_size, scoring = make_scorer(mean_absolute_error))
train_scores_mean = np.mean(train_scores, axis=1)
train_scores_std = np.std(train_scores, axis=1)
test_scores_mean = np.mean(test_scores, axis=1)
test_scores_std = np.std(test_scores, axis=1)
plt.grid()#區域
plt.fill_between(train_sizes, train_scores_mean - train_scores_std,
train_scores_mean + train_scores_std, alpha=0.1,
color="r") #填充train_mean +- train_scores_std
plt.fill_between(train_sizes, test_scores_mean - test_scores_std,
test_scores_mean + test_scores_std, alpha=0.1,
color="g")
plt.plot(train_sizes, train_scores_mean, 'o-', color='r',
label="Training score")
plt.plot(train_sizes, test_scores_mean,'o-',color="g",
label="Cross-validation score")
plt.legend(loc="best")
return plt
plot_learning_curve(LinearRegression(), 'Liner_model', train_X[:1000], train_y_ln[:1000], ylim=(0.0, 0.5), cv=5, n_jobs=1)
<module 'matplotlib.pyplot' from 'C:\\Users\\94890\\Anaconda3\\lib\\site-packages\\matplotlib\\pyplot.py'>
5.4.3 多種模型對比
train = sample_feature[continuous_feature_names + ['price']].dropna()
train_X = train[continuous_feature_names]
train_y = train['price']
train_y_ln = np.log(train_y + 1)
5.4.3 - 1 線性模型 & 嵌入式特徵選擇
本章節默認,學習者已經瞭解關於過擬合、模型複雜度、正則化等概念。否則請尋找相關資料或參考如下連接:
- 用簡單易懂的語言描述「過擬合 overfitting」? https://www.zhihu.com/question/32246256/answer/55320482
- 模型複雜度與模型的泛化能力 http://yangyingming.com/article/434/
- 正則化的直觀理解 https://blog.csdn.net/jinping_shi/article/details/52433975
在過濾式和包裹式特徵選擇方法中,特徵選擇過程與學習器訓練過程有明顯的分別。而嵌入式特徵選擇在學習器訓練過程中自動地進行特徵選擇。嵌入式選擇最常用的是L1正則化與L2正則化。在對線性迴歸模型加入兩種正則化方法後,他們分別變成了嶺迴歸與Lasso迴歸。
from sklearn.linear_model import LinearRegression
from sklearn.linear_model import Ridge
from sklearn.linear_model import Lasso
models = [LinearRegression(),
Ridge(),
Lasso()]
result = dict()
for model in models:
model_name = str(model).split('(')[0]
scores = cross_val_score(model, X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error))
result[model_name] = scores
print(model_name + ' is finished')
LinearRegression is finished
Ridge is finished
Lasso is finished
對三種方法的效果對比
result = pd.DataFrame(result)
result.index = ['cv' + str(x) for x in range(1, 6)]
result
LinearRegression | Ridge | Lasso | |
---|---|---|---|
cv1 | 0.194274 | 0.199028 | 0.392064 |
cv2 | 0.195956 | 0.200631 | 0.389369 |
cv3 | 0.195945 | 0.200816 | 0.391919 |
cv4 | 0.194693 | 0.199294 | 0.386594 |
cv5 | 0.197977 | 0.202830 | 0.392358 |
model = LinearRegression().fit(train_X, train_y_ln)
print('intercept:'+ str(model.intercept_))
sns.barplot(abs(model.coef_), continuous_feature_names)
intercept:22.490527976546055
<matplotlib.axes._subplots.AxesSubplot at 0x1834b047d88>
L2正則化在擬合過程中通常都傾向於讓權值儘可能小,最後構造一個所有參數都比較小的模型。因爲一般認爲參數值小的模型比較簡單,能適應不同的數據集,也在一定程度上避免了過擬合現象。可以設想一下對於一個線性迴歸方程,若參數很大,那麼只要數據偏移一點點,就會對結果造成很大的影響;但如果參數足夠小,數據偏移得多一點也不會對結果造成什麼影響,專業一點的說法是『抗擾動能力強』
model = Ridge().fit(train_X, train_y_ln)
print('intercept:'+ str(model.intercept_))
sns.barplot(abs(model.coef_), continuous_feature_names)
intercept:6.953548340458286
<matplotlib.axes._subplots.AxesSubplot at 0x18334871908>
L1正則化有助於生成一個稀疏權值矩陣,進而可以用於特徵選擇。如下圖,我們發現power與userd_time特徵非常重要。
model = Lasso().fit(train_X, train_y_ln)
print('intercept:'+ str(model.intercept_))
sns.barplot(abs(model.coef_), continuous_feature_names)
intercept:8.67070637212979
<matplotlib.axes._subplots.AxesSubplot at 0x183445fa988>
除此之外,決策樹通過信息熵或GINI指數選擇分裂節點時,優先選擇的分裂特徵也更加重要,這同樣是一種特徵選擇的方法。XGBoost與LightGBM模型中的model_importance指標正是基於此計算的
5.4.3 - 2 非線性模型
除了線性模型以外,還有許多我們常用的非線性模型如下,在此篇幅有限不再一一講解原理。我們選擇了部分常用模型與線性模型進行效果比對。
from sklearn.linear_model import LinearRegression
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeRegressor
from sklearn.ensemble import RandomForestRegressor
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.neural_network import MLPRegressor
from xgboost.sklearn import XGBRegressor
from lightgbm.sklearn import LGBMRegressor
models = [LinearRegression(),
DecisionTreeRegressor(),
RandomForestRegressor(),
GradientBoostingRegressor(),
MLPRegressor(solver='lbfgs', max_iter=100),
XGBRegressor(n_estimators = 100, objective='reg:squarederror'),
LGBMRegressor(n_estimators = 100)]
result = dict()
del train_X['gearbox']
for model in models:
model_name = str(model).split('(')[0]
scores = cross_val_score(model, X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error))
result[model_name] = scores
print(model_name + ' is finished')
LinearRegression is finished
DecisionTreeRegressor is finished
RandomForestRegressor is finished
GradientBoostingRegressor is finished
MLPRegressor is finished
XGBRegressor is finished
LGBMRegressor is finished
result = pd.DataFrame(result)
result.index = ['cv' + str(x) for x in range(1, 6)]
result
LinearRegression | DecisionTreeRegressor | RandomForestRegressor | GradientBoostingRegressor | MLPRegressor | XGBRegressor | LGBMRegressor | |
---|---|---|---|---|---|---|---|
cv1 | 0.194302 | 0.197484 | 0.147026 | 0.177392 | 116.598407 | 0.143866 | 0.147577 |
cv2 | 0.196005 | 0.205765 | 0.148228 | 0.179418 | 67.139661 | 0.147769 | 0.150367 |
cv3 | 0.195943 | 0.200433 | 0.149695 | 0.179679 | 26.075124 | 0.146292 | 0.148571 |
cv4 | 0.194723 | 0.197947 | 0.147885 | 0.176486 | 248.037378 | 0.144756 | 0.148429 |
cv5 | 0.197991 | 0.202071 | 0.151757 | 0.181362 | 244.320036 | 0.148483 | 0.152119 |
可以看到隨機森林模型在每一個fold中均取得了更好的效果
5.4.4 模型調參
在此我們介紹了三種常用的調參方法如下:
- 貪心算法 https://www.jianshu.com/p/ab89df9759c8
- 網格調參 https://blog.csdn.net/weixin_43172660/article/details/83032029
- 貝葉斯調參 https://blog.csdn.net/linxid/article/details/81189154
## LGB的參數集合:
objective = ['regression', 'regression_l1', 'mape', 'huber', 'fair']
num_leaves = [3,5,10,15,20,40, 55]
max_depth = [3,5,10,15,20,40, 55]
bagging_fraction = []
feature_fraction = []
drop_rate = []
5.4.4 - 1 貪心調參
best_obj = dict()
for obj in objective:
model = LGBMRegressor(objective=obj)
score = np.mean(cross_val_score(model, X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error)))
best_obj[obj] = score
best_leaves = dict()
for leaves in num_leaves:
model = LGBMRegressor(objective=min(best_obj.items(), key=lambda x:x[1])[0], num_leaves=leaves)
score = np.mean(cross_val_score(model, X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error)))
best_leaves[leaves] = score
best_depth = dict()
for depth in max_depth:
model = LGBMRegressor(objective=min(best_obj.items(), key=lambda x:x[1])[0],
num_leaves=min(best_leaves.items(), key=lambda x:x[1])[0],
max_depth=depth)
score = np.mean(cross_val_score(model, X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error)))
best_depth[depth] = score
sns.lineplot(x=['0_initial','1_turning_obj','2_turning_leaves','3_turning_depth'], y=[0.143 ,min(best_obj.values()), min(best_leaves.values()), min(best_depth.values())])
<matplotlib.axes._subplots.AxesSubplot at 0x1834d252888>
5.4.4 - 2 Grid Search 調參
from sklearn.model_selection import GridSearchCV
parameters = {'objective': objective , 'num_leaves': num_leaves, 'max_depth': max_depth}
model = LGBMRegressor()
clf = GridSearchCV(model, parameters, cv=5)
clf = clf.fit(train_X, train_y)
clf.best_params_
{'max_depth': 40, 'num_leaves': 55, 'objective': 'regression'}
model = LGBMRegressor(objective='regression',
num_leaves=55,
max_depth=40)
np.mean(cross_val_score(model, X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error)))
0.14379304572952936
5.4.4 - 3 貝葉斯調參
from bayes_opt import BayesianOptimization
def rf_cv(num_leaves, max_depth, subsample, min_child_samples):
val = cross_val_score(
LGBMRegressor(objective = 'regression_l1',
num_leaves=int(num_leaves),
max_depth=int(max_depth),
subsample = subsample,
min_child_samples = int(min_child_samples)
),
X=train_X, y=train_y_ln, verbose=0, cv = 5, scoring=make_scorer(mean_absolute_error)
).mean()
return 1 - val
rf_bo = BayesianOptimization(
rf_cv,
{
'num_leaves': (2, 100),
'max_depth': (2, 100),
'subsample': (0.1, 1),
'min_child_samples' : (2, 100)
}
)
rf_bo.maximize()
| iter | target | max_depth | min_ch... | num_le... | subsample |
-------------------------------------------------------------------------
| [0m 1 [0m | [0m 0.863 [0m | [0m 22.84 [0m | [0m 34.83 [0m | [0m 99.34 [0m | [0m 0.8796 [0m |
| [0m 2 [0m | [0m 0.8522 [0m | [0m 83.66 [0m | [0m 17.52 [0m | [0m 31.27 [0m | [0m 0.8104 [0m |
| [0m 3 [0m | [0m 0.8401 [0m | [0m 12.26 [0m | [0m 70.64 [0m | [0m 14.1 [0m | [0m 0.9035 [0m |
| [0m 4 [0m | [0m 0.8622 [0m | [0m 52.82 [0m | [0m 42.12 [0m | [0m 87.04 [0m | [0m 0.3299 [0m |
| [0m 5 [0m | [0m 0.8508 [0m | [0m 72.6 [0m | [0m 68.95 [0m | [0m 28.94 [0m | [0m 0.3273 [0m |
| [0m 6 [0m | [0m 0.863 [0m | [0m 99.96 [0m | [0m 94.19 [0m | [0m 99.23 [0m | [0m 0.5421 [0m |
| [95m 7 [0m | [95m 0.8631 [0m | [95m 90.84 [0m | [95m 2.475 [0m | [95m 99.34 [0m | [95m 0.8307 [0m |
| [0m 8 [0m | [0m 0.863 [0m | [0m 16.05 [0m | [0m 97.82 [0m | [0m 99.56 [0m | [0m 0.6616 [0m |
| [0m 9 [0m | [0m 0.8015 [0m | [0m 2.946 [0m | [0m 6.104 [0m | [0m 99.57 [0m | [0m 0.4299 [0m |
| [0m 10 [0m | [0m 0.7949 [0m | [0m 97.79 [0m | [0m 95.25 [0m | [0m 3.469 [0m | [0m 0.5979 [0m |
| [0m 11 [0m | [0m 0.8628 [0m | [0m 98.38 [0m | [0m 92.14 [0m | [0m 96.48 [0m | [0m 0.7543 [0m |
| [0m 12 [0m | [0m 0.8629 [0m | [0m 98.71 [0m | [0m 53.87 [0m | [0m 97.57 [0m | [0m 0.1395 [0m |
| [0m 13 [0m | [0m 0.7657 [0m | [0m 38.46 [0m | [0m 2.306 [0m | [0m 2.199 [0m | [0m 0.6655 [0m |
| [0m 14 [0m | [0m 0.8015 [0m | [0m 2.42 [0m | [0m 73.93 [0m | [0m 64.9 [0m | [0m 0.8561 [0m |
| [0m 15 [0m | [0m 0.7657 [0m | [0m 7.479 [0m | [0m 99.23 [0m | [0m 2.366 [0m | [0m 0.5894 [0m |
| [0m 16 [0m | [0m 0.7657 [0m | [0m 99.9 [0m | [0m 8.24 [0m | [0m 2.117 [0m | [0m 0.4889 [0m |
| [0m 17 [0m | [0m 0.863 [0m | [0m 59.63 [0m | [0m 96.77 [0m | [0m 98.46 [0m | [0m 0.6084 [0m |
| [0m 18 [0m | [0m 0.8585 [0m | [0m 99.67 [0m | [0m 35.61 [0m | [0m 55.5 [0m | [0m 0.7627 [0m |
| [0m 19 [0m | [0m 0.7657 [0m | [0m 54.1 [0m | [0m 52.56 [0m | [0m 2.834 [0m | [0m 0.7866 [0m |
| [0m 20 [0m | [0m 0.856 [0m | [0m 98.43 [0m | [0m 98.37 [0m | [0m 44.02 [0m | [0m 0.8604 [0m |
| [0m 21 [0m | [0m 0.8585 [0m | [0m 39.19 [0m | [0m 2.257 [0m | [0m 55.38 [0m | [0m 0.9486 [0m |
| [0m 22 [0m | [0m 0.8015 [0m | [0m 2.655 [0m | [0m 17.16 [0m | [0m 25.65 [0m | [0m 0.9142 [0m |
| [0m 23 [0m | [0m 0.8575 [0m | [0m 96.75 [0m | [0m 2.662 [0m | [0m 48.79 [0m | [0m 0.5475 [0m |
| [0m 24 [0m | [0m 0.8628 [0m | [0m 49.3 [0m | [0m 2.516 [0m | [0m 95.63 [0m | [0m 0.3689 [0m |
| [0m 25 [0m | [0m 0.8565 [0m | [0m 47.02 [0m | [0m 97.36 [0m | [0m 45.15 [0m | [0m 0.892 [0m |
| [95m 26 [0m | [95m 0.8632 [0m | [95m 40.13 [0m | [95m 66.03 [0m | [95m 99.89 [0m | [95m 0.1812 [0m |
| [0m 27 [0m | [0m 0.8572 [0m | [0m 42.57 [0m | [0m 48.11 [0m | [0m 48.67 [0m | [0m 0.9011 [0m |
| [0m 28 [0m | [0m 0.8597 [0m | [0m 74.9 [0m | [0m 10.61 [0m | [0m 64.42 [0m | [0m 0.9671 [0m |
| [95m 29 [0m | [95m 0.8632 [0m | [95m 66.48 [0m | [95m 28.57 [0m | [95m 99.3 [0m | [95m 0.9887 [0m |
| [0m 30 [0m | [0m 0.8592 [0m | [0m 74.97 [0m | [0m 71.13 [0m | [0m 59.46 [0m | [0m 0.9376 [0m |
=========================================================================
1 - rf_bo.max['target']
0.13680780646067503
總結
在本章中,我們完成了建模與調參的工作,並對我們的模型進行了驗證。此外,我們還採用了一些基本方法來提高預測的精度,提升如下圖所示。
plt.figure(figsize=(13,5))
sns.lineplot(x=['0_origin','1_log_transfer','2_L1_&_L2','3_change_model','4_parameter_turning'], y=[1.36 ,0.19, 0.19, 0.14, 0.13])
<matplotlib.axes._subplots.AxesSubplot at 0x1834d621708>
Task5 建模調參 END.
— By: 小雨姑娘
數據挖掘愛好者,多次獲比賽TOP名次。
作者的機器學習筆記:https://zhuanlan.zhihu.com/mlbasic
關於Datawhale:
Datawhale是一個專注於數據科學與AI領域的開源組織,彙集了衆多領域院校和知名企業的優秀學習者,聚合了一羣有開源精神和探索精神的團隊成員。Datawhale 以“for the learner,和學習者一起成長”爲願景,鼓勵真實地展現自我、開放包容、互信互助、敢於試錯和勇於擔當。同時 Datawhale 用開源的理念去探索開源內容、開源學習和開源方案,賦能人才培養,助力人才成長,建立起人與人,人與知識,人與企業和人與未來的聯結。
本次數據挖掘路徑學習,專題知識將在天池分享,詳情可關注Datawhale:
(圖片!!!)